Robert L. Jernigan

Increasing the number of related protein paralogs is important for fully understanding protein relationships, yet it remains challenging for sequences in the twilight zone. Here, we present an integrated homolog detection framework that combines sequence-based (BLASTp, MMseqs2), structure-based (Foldseek), and embedding-distance-based (PROST) similarity metrics to identify additional paralogs. To characterize functionally related protein pairs, we develop protein-family-specific supervised logistic regression models trained on curated functional annotations from MEROPS proteases and KinHub kinases. The resulting model successfully classifies proteins, with ROC-AUC of 0.99 and F1-score of 0.92 for test datasets. Applying this model, we initially identify 686 protease and 298 kinase new candidates. Subsequent structural validation, and previous annotation comparisons yield 7 new protease and 3 new kinase paralogs in the human proteome, mostly lacking prior functional characterization. An additional outcome is structural identification of catalytically important residues for larger numbers of proteases and kinases. Despite the small number of new paralogs for well-studied proteases and kinases, our results demonstrate that integrating orthogonal homolog approaches with family-specific regression models provides a robust, scalable strategy for discovering new functionally related proteins, which is a generalizable approach for novel protein function discovery and can be applied more broadly to under-annotated proteomes.

Multiple sequence alignment (MSA) data play a crucial role in the study of protein mutations, with contact prediction being a notable application. Existing methods are often model-based or algorithmic and typically do not incorporate statistical inference to quantify the uncertainty of the prediction outcomes. To address this, we propose a novel framework that transforms the task of contact prediction into a statistical testing problem. Our approach is motivated by the partial correlation for continuous random variables. With one-hot encoding of MSA data, we are able to construct a partial correlation graph for multivariate categorical variables. In this framework, two connected nodes in the graph indicate that the corresponding positions on the protein form a contact. A new spectrum-based test statistic is introduced to test whether two positions are partially correlated. Moreover, the new framework enables the identification of amino acid combinations that contribute to the correlation within the identified contacts, an important but largely unexplored aspect of protein mutations. Numerical experiments demonstrate that our proposed method is valid in terms of controlling Type I errors and powerful in general. Real data applications on various protein families further validate the practical utility of our approach in coevolution and mutation analysis.

The identification of paralogs is critical for understanding protein evolution, function, and for drug design, yet many human proteins remain unannotated and poorly classified. Sequence-based homology detection alone often fails to detect distant paralogs, especially in the “twilight zone” or beyond, regarding sequence identity. Here we present an integrated homolog detection framework that combines results from BLASTp, MMseqs2, Foldseek, and the large protein language model-based tool PROST, followed by validation based on comparison of structures and for enzymes comparison of the specific structures of the catalytic residues. Using all-versus-all exhaustive comparisons across the 20,647 human proteins, we systematically identify novel paralogs and assess their catalytic residues for two serine protease clans. We discovered 14 previously uncharacterized human serine carboxypeptidases, validated against experimentally determined PDB structures, with 11 of these displaying conserved catalytic triads. We further identify 203 new paralogs for human kinases, with 163 of these in the major clusters that represent previously uncharacterized kinase subtypes and 30 putative novel human transcription factors. Across both serine protease subtypes, structural alignments enable the prediction of the previously unknown catalytic residues for those lacking UniProt annotations of active site residues. By integrating sequence, structure, and LPLM embedding-based approaches, the framework enables the discovery of surprisingly large numbers of unknown paralogs, permitting defining catalytic residues, and expands the understanding of protein functional landscapes. These findings provide the foundation for a large number of future functional, evolutionary, and therapeutic investigations.

Accurately predicting protein thermostability is crucial for numerous applications in biotechnology, pharmaceuticals, and food science. Experimental methods for determining protein melting temperatures are often time-consuming and costly, driving the need for efficient computational alternatives. In this paper, we introduce ESMStabP, an enhanced regression model for predicting protein thermostability. To improve model performance and generalizability, we assembled a significantly larger dataset by combining and cleaning datasets previously utilized in other thermostability models. Building on Deep-StabP, ESMStabP incorporates significant improvements, using embeddings from the ESM2 protein language model and thermophilic classifications. The predictions from ESMStabP outperform DeepStabP and other existing predictors, achieving an R ² of 0.95 and a Pearson correlation coefficient (PCC) of 0.97. Despite these improvements, challenges such as dataset availability. This work underscores the critical role of specific layer identification for model development and outlines potential directions for future advancements in protein stability predictions

The JCVI-Syn3 organism is a minimal organism derived from Mycoplasma mycoides capri, which is capable of self-replication. While the ancestor has 863 genes, the synthetic progeny has only 473, with 434 of these coding for proteins. Despite initial efforts to understand all functions of the organism, a significant number of these protein-coding genes still have unknown functions, and subsequent studies have been only partially successful in elucidating their roles. In this study, we employ our innovative method PROST to identify homologs and better understand these previously unidentified genes. PROST employs protein language embeddings and enables the identification of remote homologs with as low as 16% sequence identity. PROST successfully finds functionally annotated homologs for 93% of the minimal genome with a high level of accuracy, both confirming previously identified functions, as well as proposing new functions for others. The results of our study can be accessed at https://bit.ly/prost-syn3a.